Neural network pruning for face anti-spoofing

Abstract

With the development of smartphones and mobile payments, face recognition systems have been rapidly deployed and face anti-spoofing research has become increasingly popular due to many attacks which are threatening the security of users' data. The generalization performance of the face anti-spoofing(FAS) model is worth doing research due to the threats from unseen and different types from trained types of attacks. The popularity of wearable devices has more requirements: smaller volume, fewer parameters, and faster inference speed, for the FAS model due to the limitation of device computing power and volume. Model pruning is a very popular way of model compression, but the previous model pruning did not pay attention to the generalization ability of the pruned model. Therefore, we explore the effect of model pruning on the generalization capability of face anti-spoofing models and propose Meta-MMD based on MetaPruning and Maximum Mean Discrepancy(MMD) for Face anti-spoofing. After testing on the mainstream FAS datasets, the generalization performance increases after Meta-MMD pruning compared with using Metapruning.